Seeking Deep Thoughts

Thoughts that came to mind as I’m reading the Deepseek Maths paper

I started reading the Deepseek Math paper, after recently finishing the R1 paper. The following thoughts started coming to mind:

• R1 is successful simply because of the training process
• Deepseek Math is worked really well because of data quality
• GRPO is used only to reduce memory usage (still reading the Deepseek Math paper; I could be wrong)

Takeaway is that a lot can be done by simply flipping and rearranging all the existing levers and switches we already have.